Observer

Watch a hard video with a coding agent sitting next to you.

Paste a podcast, a lecture, or a conference talk. Observer transcribes it, serves a page with the player, and hands your CLI agent a set of MCP tools: read the transcript, prepare the concepts, build the charts, then answer the moment you pause. One process serves the page and speaks MCP, so there is nothing to start by hand.

How a session goes

1. Feed. Paste the URL, in the page or to the agent. Say whether the video carries ads, and whether you want the research pass and the visual toolkit.
2. Transcript. Captions when the video has them, speech recognition when it does not.
3. Research (optional). The agent searches the concepts it did not already know, and looks specifically for what moved after the recording date, so a two-year-old talk arrives with its corrections attached.
4. Toolkit (optional). The agent writes the charts and diagrams for the equations and systems in the talk, compiles them, and fixes them until they render.
5. Session. You watch. You pause. You ask, by voice or by typing. The answer comes back in text first, spoken if you want it, with a prepared visual behind it when one fits.

The preparation is the point: by the time you press play, the transcript, the definitions, the jargon list, the research and the visuals are already in the agent's context. A doubt at minute 34 gets an answer, not a wait.

What the agent knows when you pause

Your position in the video is the join key. Every pause tells the agent the second you stopped at, the words around it, and which prepared concepts cover that moment. It answers about what was actually being said, not about the video in general.

Install

Needs Node 24.12 or newer. yt-dlp widens which videos have captions, and ffmpeg is needed only for the speech-recognition path.

git clone https://github.com/hec-ovi/observer-skill
cd observer-skill
npm install
node bin/observer.ts doctor

doctor reports what your machine can do and names what to install to widen it.

Point a CLI agent at it:

claude mcp add observer -- node /path/to/observer-skill/bin/observer.ts mcp

Then tell it to study something, or run node bin/observer.ts serve and paste a URL into the page yourself.

Tools

Tool	Does
open	Take a URL and options, start transcription, open the page
status	Phase, progress, counts, whether a page is connected, what is still missing
transcript	Read the transcript whole or by time range, paginated
concepts	Write the concept list: definitions, equations, systems, jargon
note	Pin a research finding to a concept
build	Compile a visual, verify it in the real page, return a snapshot to look at
link	Bind a visual to a concept and a stretch of video
ready	Close preparation, unlock the player
wait	Wait for the next pause, question, or setting change
where	The same context on demand, for a question typed in the terminal
say	Answer, optionally spoken, optionally showing a visual
show / hide	Move the stage between video and visual

Phases, error set and the full argument list: docs/ARCHITECTURE.md.

Visuals

A visual is a small ES module that mounts into the stage and gets the current theme. It can use ECharts for interactive and large series, D3 for bespoke geometry and simulations, and KaTeX for equations, all shared through one import map so the tenth chart costs nothing to open.

Before anything reaches your screen it is bundled, linted, and mounted in a hidden sandbox inside the page you already have open, under a content policy derived from that document. The agent sees the real errors and a snapshot of the result, and iterates until it is right. A visual that has not passed cannot be shown.

Voice

Speaking defaults to the browser's own voice, which costs nothing to load. Kyutai Pocket TTS runs entirely in the browser through ONNX Runtime Web for a better one: the settings panel states the one-time download before you agree to it, shows it arriving, and credits Kyutai while it is the chosen voice. Listening is hold to talk. Both sit behind a port with three providers each (browser, local model, OpenAI-compatible endpoint), so moving either to a cloud voice is a setting and not a rewrite.

Design

Layered surfaces, soft corners, one accent on the thing you are meant to press, one title per thing, and nothing on screen that is not carrying signal. Light, dark and system reach the charts too: the theme travels into every visual, so switching restyles them in place without rebuilding. Transitions between the video and a visual are short and respect reduced motion.

Configuration

Environment variables, all with working defaults. node bin/observer.ts --help lists them with the values in force on your machine.

Variable	Default	What
OBSERVER_PORT	4830	First port to try for the page; a taken one moves up
OBSERVER_HOME	$XDG_DATA_HOME/observer	Where sessions, transcripts and artifacts live
OBSERVER_TRANSCRIPT	auto	auto, captions, endpoint-asr, or file
OBSERVER_ASR_URL	unset	OpenAI-compatible transcription endpoint
YTDLP_BIN	yt-dlp	The caption fetcher binary, when it is not on PATH
FFMPEG_BIN	ffmpeg	The audio converter, when it is not on PATH

Repository

Path	What
boxes/	One folder per box, each with its CONTRACT.md
boxes/agent-io/prompts/	The prompts the agent works from
skills/observer/	The skill the agent loads
docs/	Index, architecture, decisions, plan
bin/	The CLI: mcp, serve, doctor

Every box is built against its contract alone and tested through its real entry point. node --run test runs all of them; node --run typecheck covers both TypeScript projects.

MIT.